Predicting drug-disease interactions is essential for pharmacological discovery, yet most computational models underperform in cold-start scenarios where entities have limited interaction history. Standard Graph Neural Networks (GNNs) typically require dense connectivity, leaving sparse nodes as topological informationally isolated within biological networks. This study addresses such sparsity by constructing a unified heterogeneous knowledge graph using drug-gene (ChG-Miner) and drug-disease (DCh-Miner) associations from the BioSNAP database. We demonstrate that gene-target interactions serve as a latent bridge, allowing the model to infer context for drugs with minimal disease history through their genomic profiles. We evaluate two GNN architectures — GraphSAGE and Graph Attention Networks (GAT) — under severe sparsity, where over 50% of nodes exhibit a degree d ≤ 1.Results reveal a significant architectural trade-off: while GAT is effective in data-rich regions, its performance drops to a ROC-AUC of 0.79 for degree-1 nodes. In contrast, GraphSAGE maintains a ROC-AUC of 0.97 in the same sparse regions. This suggests that attention mechanisms become a liability when local neighborhoods are insufficient for statistical weighting. Our findings indicate that while attention-based models excel in well-annotated contexts, mean-pooling aggregators like GraphSAGE provide superior robustness for rare diseases and novel compound association prediction.